File size: 7,143 Bytes
970a7a2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 |
import bisect
import copy
import math
from collections import defaultdict
from itertools import repeat, chain
import numpy as np
import torch
import torch.utils.data
import torchvision
from PIL import Image
from torch.utils.data.sampler import BatchSampler, Sampler
from torch.utils.model_zoo import tqdm
def _repeat_to_at_least(iterable, n):
repeat_times = math.ceil(n / len(iterable))
repeated = chain.from_iterable(repeat(iterable, repeat_times))
return list(repeated)
class GroupedBatchSampler(BatchSampler):
"""
Wraps another sampler to yield a mini-batch of indices.
It enforces that the batch only contain elements from the same group.
It also tries to provide mini-batches which follows an ordering which is
as close as possible to the ordering from the original sampler.
Args:
sampler (Sampler): Base sampler.
group_ids (list[int]): If the sampler produces indices in range [0, N),
`group_ids` must be a list of `N` ints which contains the group id of each sample.
The group ids must be a continuous set of integers starting from
0, i.e. they must be in the range [0, num_groups).
batch_size (int): Size of mini-batch.
"""
def __init__(self, sampler, group_ids, batch_size):
if not isinstance(sampler, Sampler):
raise ValueError(f"sampler should be an instance of torch.utils.data.Sampler, but got sampler={sampler}")
self.sampler = sampler
self.group_ids = group_ids
self.batch_size = batch_size
def __iter__(self):
buffer_per_group = defaultdict(list)
samples_per_group = defaultdict(list)
num_batches = 0
for idx in self.sampler:
group_id = self.group_ids[idx]
buffer_per_group[group_id].append(idx)
samples_per_group[group_id].append(idx)
if len(buffer_per_group[group_id]) == self.batch_size:
yield buffer_per_group[group_id]
num_batches += 1
del buffer_per_group[group_id]
assert len(buffer_per_group[group_id]) < self.batch_size
# now we have run out of elements that satisfy
# the group criteria, let's return the remaining
# elements so that the size of the sampler is
# deterministic
expected_num_batches = len(self)
num_remaining = expected_num_batches - num_batches
if num_remaining > 0:
# for the remaining batches, take first the buffers with largest number
# of elements
for group_id, _ in sorted(buffer_per_group.items(), key=lambda x: len(x[1]), reverse=True):
remaining = self.batch_size - len(buffer_per_group[group_id])
samples_from_group_id = _repeat_to_at_least(samples_per_group[group_id], remaining)
buffer_per_group[group_id].extend(samples_from_group_id[:remaining])
assert len(buffer_per_group[group_id]) == self.batch_size
yield buffer_per_group[group_id]
num_remaining -= 1
if num_remaining == 0:
break
assert num_remaining == 0
def __len__(self):
return len(self.sampler) // self.batch_size
def _compute_aspect_ratios_slow(dataset, indices=None):
print(
"Your dataset doesn't support the fast path for "
"computing the aspect ratios, so will iterate over "
"the full dataset and load every image instead. "
"This might take some time..."
)
if indices is None:
indices = range(len(dataset))
class SubsetSampler(Sampler):
def __init__(self, indices):
self.indices = indices
def __iter__(self):
return iter(self.indices)
def __len__(self):
return len(self.indices)
sampler = SubsetSampler(indices)
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=1,
sampler=sampler,
num_workers=14, # you might want to increase it for faster processing
collate_fn=lambda x: x[0],
)
aspect_ratios = []
with tqdm(total=len(dataset)) as pbar:
for _i, (img, _) in enumerate(data_loader):
pbar.update(1)
height, width = img.shape[-2:]
aspect_ratio = float(width) / float(height)
aspect_ratios.append(aspect_ratio)
return aspect_ratios
def _compute_aspect_ratios_custom_dataset(dataset, indices=None):
if indices is None:
indices = range(len(dataset))
aspect_ratios = []
for i in indices:
height, width = dataset.get_height_and_width(i)
aspect_ratio = float(width) / float(height)
aspect_ratios.append(aspect_ratio)
return aspect_ratios
def _compute_aspect_ratios_coco_dataset(dataset, indices=None):
if indices is None:
indices = range(len(dataset))
aspect_ratios = []
for i in indices:
img_info = dataset.coco.imgs[dataset.ids[i]]
aspect_ratio = float(img_info["width"]) / float(img_info["height"])
aspect_ratios.append(aspect_ratio)
return aspect_ratios
def _compute_aspect_ratios_voc_dataset(dataset, indices=None):
if indices is None:
indices = range(len(dataset))
aspect_ratios = []
for i in indices:
# this doesn't load the data into memory, because PIL loads it lazily
width, height = Image.open(dataset.images[i]).size
aspect_ratio = float(width) / float(height)
aspect_ratios.append(aspect_ratio)
return aspect_ratios
def _compute_aspect_ratios_subset_dataset(dataset, indices=None):
if indices is None:
indices = range(len(dataset))
ds_indices = [dataset.indices[i] for i in indices]
return compute_aspect_ratios(dataset.dataset, ds_indices)
def compute_aspect_ratios(dataset, indices=None):
if hasattr(dataset, "get_height_and_width"):
return _compute_aspect_ratios_custom_dataset(dataset, indices)
if isinstance(dataset, torchvision.datasets.CocoDetection):
return _compute_aspect_ratios_coco_dataset(dataset, indices)
if isinstance(dataset, torchvision.datasets.VOCDetection):
return _compute_aspect_ratios_voc_dataset(dataset, indices)
if isinstance(dataset, torch.utils.data.Subset):
return _compute_aspect_ratios_subset_dataset(dataset, indices)
# slow path
return _compute_aspect_ratios_slow(dataset, indices)
def _quantize(x, bins):
bins = copy.deepcopy(bins)
bins = sorted(bins)
quantized = list(map(lambda y: bisect.bisect_right(bins, y), x))
return quantized
def create_aspect_ratio_groups(dataset, k=0):
aspect_ratios = compute_aspect_ratios(dataset)
bins = (2 ** np.linspace(-1, 1, 2 * k + 1)).tolist() if k > 0 else [1.0]
groups = _quantize(aspect_ratios, bins)
# count number of elements per group
counts = np.unique(groups, return_counts=True)[1]
fbins = [0] + bins + [np.inf]
print(f"Using {fbins} as bins for aspect ratio quantization")
print(f"Count of instances per bin: {counts}")
return groups
|